Learning to Unscramble with Agentic AI: A New Self-Supervised Approach to Symbolic Simplification

by Prof. David Shih (Rutgers)

Friday 12 Jun 2026, 11:00 → 13:00 US/Eastern

CFNS Library (https://bnl.zoomgov.com/j/1614715193?pwd=WkwxODVWdzZzb29zQnZRVGp3VTBDQT09)

Abstract: I will give an in-depth overview of my two recent papers [2603.11164 and 2604.05034] on a highly effective new ML method for symbolic simplification of mathematical expressions based on a Markov Decision Process trained with self-supervised learning. Training data is generated automatically by scrambling simple expressions into complicated ones using a set of mathematical identities, and recording the reverse unscrambling steps. When applied to dilog sums and tree-level YM scattering amplitudes, our method is shown to have nearly perfect simplification rate, far surpassing previous approaches based on reinforcement learning and end-to-end regression. When extended to IBP reduction, we show (in the toy 3pt two-loop triangle-box setting) that the method is able to reduce integrals of increasing weight while remaining flat in memory consumption, compared to general purpose Laporta-based algorithms such as Kira where the memory consumption rises rapidly with weight. These projects were both done in collaboration with Claude Code, and I will conclude with some lessons learned and general thoughts for the future of AI and our field.